Leveraging diffusion models for unsupervised out-of-distribution detection on image manifold

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Reliability of machine learning models is challenged by out-of-distribution (OOD) data.
• Images are hypothesized to exist on manifolds defined by latent properties like color, position, and shape.

Purpose of the Study:

• To propose a novel diffusion model-based approach for robust OOD detection in images.
• To leverage the manifold hypothesis for distinguishing in-domain from OOD images.

Main Methods:

• Training a diffusion model on in-domain images.
• Employing a masking process to lift images from their original manifold.
• Mapping lifted images towards the in-domain manifold using the diffusion model.
• Quantifying OOD detection by measuring the distance between original and mapped images.

Main Results:

• The proposed method demonstrated strong and consistent performance across diverse datasets with variations in color, semantics, and resolution.
• Effective detection of OOD images was achieved, confirming the approach's versatility.
• Ablation studies validated the importance of individual framework components.

Conclusions:

• The diffusion model approach offers an effective strategy for OOD detection in image analysis.
• The method's ability to handle diverse image characteristics underscores its practical applicability.
• This work contributes to enhancing the reliability of machine learning models in real-world scenarios.